sky.py

   1 #!/usr/bin/env python
   2
   3 # Any copyright is dedicated to the Public Domain.
   4 # https://creativecommons.org/publicdomain/zero/1.0/
   5
   6 # Written by Francois Fleuret <francois@fleuret.org>
   7
   8 import math, sys, tqdm, os
   9
  10 import torch, torchvision
  11
  12 from torch import nn
  13 from torch.nn import functional as F
  14
  15 ######################################################################
  16
  17 import problem
  18
  19
  20 class Sky(problem.Problem):
  21     colors = torch.tensor(
  22         [
  23             [255, 255, 255],
  24             [255, 0, 0],
  25             [0, 192, 0],
  26             [0, 0, 255],
  27             [255, 192, 0],
  28             [0, 255, 255],
  29             [255, 0, 255],
  30             [192, 255, 192],
  31             [255, 192, 192],
  32             [192, 192, 255],
  33             [192, 192, 192],
  34         ]
  35     )
  36
  37     token_background = 0
  38     first_bird_token = 1
  39     nb_bird_tokens = colors.size(0) - 1
  40     token_forward = first_bird_token + nb_bird_tokens
  41     token_backward = token_forward + 1
  42
  43     token2char = (
  44         "_" + "".join([chr(ord("A") + n) for n in range(len(colors) - 1)]) + "><"
  45     )
  46
  47     def __init__(self, height=6, width=8, nb_birds=3, speed=1, nb_iterations=4):
  48         self.height = height
  49         self.width = width
  50         self.nb_birds = nb_birds
  51         self.speed = speed
  52         self.nb_iterations = nb_iterations
  53
  54     def direction_tokens(self):
  55         return self.token_forward, self.token_backward
  56
  57     def generate_frame_sequences(self, nb):
  58         frame_sequences = []
  59
  60         for _ in tqdm.tqdm(range(nb), dynamic_ncols=True, desc="world generation"):
  61             result = torch.zeros(
  62                 self.nb_iterations, self.height, self.width, dtype=torch.int64
  63             )
  64
  65             i, j, vi, vj = (
  66                 torch.empty(self.nb_birds, dtype=torch.int64),
  67                 torch.empty(self.nb_birds, dtype=torch.int64),
  68                 torch.empty(self.nb_birds, dtype=torch.int64),
  69                 torch.empty(self.nb_birds, dtype=torch.int64),
  70             )
  71
  72             col = (
  73                 torch.randperm(self.colors.size(0) - 1)[: self.nb_birds].sort().values
  74                 + 1
  75             )
  76
  77             for n in range(self.nb_birds):
  78                 while True:
  79                     i[n] = torch.randint(self.height, (1,))
  80                     j[n] = torch.randint(self.width, (1,))
  81                     vm = torch.randint(4, (1,))
  82                     vi[n], vj[n] = (vm % 2) * 2 - 1, (vm // 2) * 2 - 1
  83                     if (
  84                         i[n] - vi[n] >= 0
  85                         and i[n] - vi[n] < self.height
  86                         and j[n] - vj[n] >= 0
  87                         and j[n] - vj[n] < self.width
  88                     ):
  89                         break
  90
  91             for l in range(self.nb_iterations):
  92                 for n in range(self.nb_birds):
  93                     c = col[n]
  94                     result[l, i[n], j[n]] = c
  95                     result[l, i[n] - vi[n], j[n]] = c
  96                     result[l, i[n], j[n] - vj[n]] = c
  97
  98                     if (i[n] == 0 and vi[n] == -1) or (
  99                         i[n] == self.height - 1 and vi[n] == 1
 100                     ):
 101                         vi[n] = -vi[n]
 102
 103                     if (j[n] == 0 and vj[n] == -1) or (
 104                         j[n] == self.width - 1 and vj[n] == 1
 105                     ):
 106                         vj[n] = -vj[n]
 107
 108                     i[n] += vi[n]
 109                     j[n] += vj[n]
 110
 111             frame_sequences.append(result)
 112
 113         return frame_sequences
 114
 115     def generate_token_sequences(self, nb):
 116         frame_sequences = self.generate_frame_sequences(nb)
 117
 118         result = []
 119
 120         for frame_sequence in frame_sequences:
 121             a = []
 122             if torch.rand(1) < 0.5:
 123                 for frame in frame_sequence:
 124                     if len(a) > 0:
 125                         a.append(torch.tensor([self.token_forward]))
 126                     a.append(frame.flatten())
 127             else:
 128                 for frame in reversed(frame_sequence):
 129                     if len(a) > 0:
 130                         a.append(torch.tensor([self.token_backward]))
 131                     a.append(frame.flatten())
 132
 133             result.append(torch.cat(a, dim=0)[None, :])
 134
 135         return torch.cat(result, dim=0)
 136
 137     ######################################################################
 138
 139     def frame2img(self, x, scale=15):
 140         x = x.reshape(-1, self.height, self.width)
 141         m = torch.logical_and(
 142             x >= 0, x < self.first_bird_token + self.nb_bird_tokens
 143         ).long()
 144         x = self.colors[x * m].permute(0, 3, 1, 2)
 145         s = x.shape
 146         x = x[:, :, :, None, :, None].expand(-1, -1, -1, scale, -1, scale)
 147         x = x.reshape(s[0], s[1], s[2] * scale, s[3] * scale)
 148
 149         x[:, :, :, torch.arange(0, x.size(3), scale)] = 0
 150         x[:, :, torch.arange(0, x.size(2), scale), :] = 0
 151         x = x[:, :, 1:, 1:]
 152
 153         for n in range(m.size(0)):
 154             for i in range(m.size(1)):
 155                 for j in range(m.size(2)):
 156                     if m[n, i, j] == 0:
 157                         for k in range(2, scale - 2):
 158                             for l in [0, 1]:
 159                                 x[n, :, i * scale + k, j * scale + k - l] = 0
 160                                 x[
 161                                     n, :, i * scale + scale - 1 - k, j * scale + k - l
 162                                 ] = 0
 163
 164         return x
 165
 166     def seq2img(self, seq, scale=15):
 167         all = [
 168             self.frame2img(
 169                 seq[:, : self.height * self.width].reshape(-1, self.height, self.width),
 170                 scale,
 171             )
 172         ]
 173
 174         separator = torch.full((seq.size(0), 3, self.height * scale - 1, 1), 0)
 175
 176         t = self.height * self.width
 177
 178         while t < seq.size(1):
 179             direction_tokens = seq[:, t]
 180             t += 1
 181
 182             direction_images = self.colors[
 183                 torch.full(
 184                     (direction_tokens.size(0), self.height * scale - 1, scale), 0
 185                 )
 186             ].permute(0, 3, 1, 2)
 187
 188             for n in range(direction_tokens.size(0)):
 189                 if direction_tokens[n] == self.token_forward:
 190                     for k in range(scale):
 191                         for l in [0, 1]:
 192                             direction_images[
 193                                 n,
 194                                 :,
 195                                 (self.height * scale) // 2 - scale // 2 + k - l,
 196                                 3 + scale // 2 - abs(k - scale // 2),
 197                             ] = 0
 198                 elif direction_tokens[n] == self.token_backward:
 199                     for k in range(scale):
 200                         for l in [0, 1]:
 201                             direction_images[
 202                                 n,
 203                                 :,
 204                                 (self.height * scale) // 2 - scale // 2 + k - l,
 205                                 3 + abs(k - scale // 2),
 206                             ] = 0
 207                 else:
 208                     for k in range(2, scale - 2):
 209                         for l in [0, 1]:
 210                             direction_images[
 211                                 n,
 212                                 :,
 213                                 (self.height * scale) // 2 - scale // 2 + k - l,
 214                                 k,
 215                             ] = 0
 216                             direction_images[
 217                                 n,
 218                                 :,
 219                                 (self.height * scale) // 2 - scale // 2 + k - l,
 220                                 scale - 1 - k,
 221                             ] = 0
 222
 223             all += [
 224                 separator,
 225                 direction_images,
 226                 separator,
 227                 self.frame2img(
 228                     seq[:, t : t + self.height * self.width].reshape(
 229                         -1, self.height, self.width
 230                     ),
 231                     scale,
 232                 ),
 233             ]
 234
 235             t += self.height * self.width
 236
 237         return torch.cat(all, dim=3)
 238
 239     def seq2str(self, seq):
 240         result = []
 241         for s in seq:
 242             result.append("".join([self.token2char[v] for v in s]))
 243         return result
 244
 245     def save_image(self, input, result_dir, filename):
 246         img = self.seq2img(input.to("cpu"))
 247         image_name = os.path.join(result_dir, filename)
 248         torchvision.utils.save_image(img.float() / 255.0, image_name, nrow=6, padding=4)
 249
 250     def save_quizzes(self, input, result_dir, filename_prefix):
 251         self.save_image(input, result_dir, filename_prefix + ".png")
 252
 253
 254 ######################################################################
 255
 256 if __name__ == "__main__":
 257     import time
 258
 259     sky = Sky(height=6, width=8, speed=1, nb_iterations=4)
 260
 261     start_time = time.perf_counter()
 262     seq = sky.generate_frame_sequences(nb=64)
 263     delay = time.perf_counter() - start_time
 264     print(f"{seq.size(0)/delay:02f} seq/s")
 265
 266     # print(sky.seq2str(seq[:4]))
 267
 268     # for t in range(len(it[0])):
 269     # img = torch.cat([sky.frame2img(f[t]) for f in it], dim=0)
 270     # torchvision.utils.save_image(
 271     # img.float() / 255.0,
 272     # f"/tmp/frame_{t:03d}.png",
 273     # nrow=8,
 274     # padding=6,
 275     # pad_value=0,
 276     # )
 277
 278     # m = (torch.rand(seq.size()) < 0.05).long()
 279     # seq = (1 - m) * seq + m * 23
 280
 281     print(seq.size())
 282     img = sky.seq2img(seq)
 283     print(img.size())
 284
 285     torchvision.utils.save_image(
 286         img.float() / 255.0, "/tmp/world.png", nrow=6, padding=6, pad_value=0
 287     )